Data heterogeneity and algorithmic bias in AI-based antimicrobial resistance prediction: a systematic review and mitigation framework

Abstract

Antimicrobial resistance (AMR) represents one of the most critical global public health threats of the contemporary era, contributing to millions of deaths and substantial morbidity worldwide. Artificial intelligence and machine learning (AI/ML) have been increasingly applied to AMR detection, prediction, and management, demonstrating promising results in controlled research settings. However, clinical translation remains substantially constrained by fundamental methodological challenges, particularly data heterogeneity and algorithmic bias, whose extent and consequences are not yet adequately characterised. A systematic review with descriptive synthesis was conducted in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA 2020) guidelines. Searches were performed across PubMed/MEDLINE, Embase, Scopus, Web of Science, IEEE Xplore, and the ACM Digital Library, covering studies published between January 2020 and April 2026. Eligible studies applied AI/ML approaches to AMR prediction using whole-genome sequencing (WGS), antimicrobial susceptibility testing (AST), electronic health records (EHRs), surveillance datasets, and spectral data. Data were extracted using a standardized charting form and synthesized using descriptive and thematic methods. No meta-analysis was conducted owing to substantial methodological heterogeneity across included studies. A total of 68 studies published between 2020 and 2026 were included, of which 21 underwent detailed analysis. Studies demonstrated substantial heterogeneity across data modalities, patient populations, laboratory practices, and geographic settings. WGS-based approaches were most frequently represented, followed by AST- and EHR-based models. While many models achieved strong internal performance, generalizability across external settings remained limited. Key sources of heterogeneity included variability in data modalities, laboratory protocols, population composition, and temporal and geographic distribution. Major forms of bias identified included sampling bias, label inconsistency, structural confounding, and clinical context bias. Mitigation strategies demonstrated partial and context-dependent improvements. Data heterogeneity and algorithmic bias are the primary constraints on AI/ML-based AMR prediction, not algorithmic sophistication. The strong internal performance metrics reported across the literature do not reliably generalize to diverse populations, geographic settings, or clinical environments. Meaningful progress requires a deliberate shift in research priorities toward globally representative datasets, harmonized laboratory standards, mandatory external validation, and equitable model development. The proposed Heterogeneity Mitigation Framework offers a structured, evidence-grounded approach to these challenges; its empirical validation in diverse real-world settings is the most important next step for the field. Although this review was not prospectively registered in PROSPERO, it was conducted using predefined eligibility criteria, structured data extraction, and transparent reporting in accordance with PRISMA 2020 guidelines.